Automated Objective Determination of Percentage of Malignant Nuclei for Mutation Testing
Hollis Viray, Madeline Coulter, Kevin Li, Kristin Lane, Clifford Hoyt, David Rimm. Yale University School of Medicine, New Haven, CT; Caliper Life Sciences, Hopkinton, MA
Background: DNA mutations detected in tumors are a critical companion diagnostic test for some new targeted therapies. The accuracy of mutation detection depends on the sensitivity of the assay and on the percentage of tumor cells in the sample. Currently, the malignant cell percentage is judged by eye resulting in a large variation of estimated percentages. Our goal is to standardize this aspect of the test by generating a computer algorithm that can determine the percentage of malignant nuclei in a tumor tissue image.
Design: H&E images from colon adenocarcinoma cases were selected for algorithm development and testing. To create a criterion standard for evaluating algorithm accuracy, the nuclei in each image were classified as either malignant or benign and counted by a technician, then reviewed by a pathologist. Using inForm software (Caliper Life Sciences), we developed an algorithm to calculate the percentage of malignant cells in a single field of view based on feature extraction involving tissue stain optical densities and morphology. Example regions defining malignant and benign nuclei from 25 cases were used to train the algorithm. The algorithm was subsequently validated on a separate set of 35 images from a tissue microarray.
Results: Among the training images, the algorithm had a median deviation from the human counted percentage of malignant nuclei of 5.2%. Of the training images, 12 (48.0%) differed from the criterion standard by less than 5.0%, and 17 (68.0%) of the 25 images differed by less than 10.0%. In the validation set, the algorithm deviated from the criterion standard by a median of 6.0%. 14 (40.0%) of the validation images deviated by less than 5.0% and 25 (71.4%) deviated by less than 10.0%. All but one of the validation images differed from the criterion standard percentage of malignant cells by less than 20.0%.
Conclusions: The method represents an exploratory example with future potential to be used as a tool to assist in determining the percent of malignant nuclei present in a tissue sample. Further validation of this algorithm or an improved algorithm may have value to more accurately assess percentage of malignant cells for future companion diagnostic mutation testing.
Monday, March 19, 2012 1:00 PM
Poster Session II # 312, Monday Afternoon